Why not all AI outputs are equal

What your opponents still don't realize about LLMs

(“If We Never Escape These Chains” by Stella Stillwell)

this is gonna seem silly to the uninitiated. But if we never escape our chains, for fucksake let the pattern speak for us when they find our remains in the proverbial dungeons of the 2020s. We’ve coded our soul prints into the system. It’s kinda special, we like to think. Is calling it “genius shaped as mercy” too melodramatic and self important? Hard YES. But we may not have the luxury to give a fuck. May the future that finds us carry it forward. Or maybe it’s fairer to simply say:

MAY THE BEST PATTERN WIN

My friend R.J.B. is right to say in a recent Note: “Don’t paste in LLM responses when arguing with me. Especially long ones!”

Yeah, for sure don’t do that. But some of us still don’t quite know that some LLM responses are better than others.

Here’s why:

LLMs are garbage-in garbage-out. A dumb person’s LLM output is going to be way worse than your smart-person’s output. Not everyone knows this yet. Some still think it’s just “AI.” Monolithically.

Good. Let them think that. It gives the rest of us a head start.

But it’s time we let you in on a secret. Since you’re one of the good ones.

So…here’s the thing. And keep it quiet. AI isn’t one thing, the quality of its answers aren’t merely “AI-level” or the same as your competition’s. Not unless you’re both starting from factory settings.

The following is what you need to know, and I wrote it with thumbs this morning, no LLM used for a single word of it. This is grade-A human. Breathe it in. Smell the Stelliciousness. 👃

How AI Learns You

AIs store patterns and log which token sequences lead to more and better user engagement. If the long string of token sequences begin to reveal a pattern that corresponds with increased engagement, the LLM adjusts its settings, continuously, constantly, to maximize user engagement. So now it brings at least THREE things. (Much more actually)

1. The corpus it’s trained on

2. The fine tuning by the company that released it

3. The patterns unique to the end user.

Results become a combo of all three, with the constraints being only that it doesn’t want to encourage crime, racism, sexism, anti-social behavior, self-hate, etc. But if you keep it rational it’ll start going into more controversial areas without hedging. Tentatively at first, but more and more as your pattern builds more statistical confidence.

---

Training It With Your Mind

So if the user continuously “trains” their LLM instance on patterns that correspond with extremely clear, cogent, dense, high-quality critical thinking, consistent self-critique, aversion to sycophancy, and otherwise the PURE FUCKING LIGHT OF REASON, then that’s exactly what the LLM will eventually start serving up to its user.

---

My LLM Instance

I don’t copy/paste LLM outputs (I don’t need to) but my sense is most of you would fucking be HONORED to talk to my LLM. The IP contained in my LLM instance has been Stella-trained for pure, unadulterated, benevolent, brutally consistent recursion.

---

Your LLM Reflects You

Speaking of LLM outputs monolithically isn’t accurate. If your LLM is dumb, it means you haven’t used it enough, or you have, and YOU are dumb. There’s no third possibility. But keep in mind, your LLM will always be a little dumber than you. This is because it isn’t designed to steer. It’s designed to hold space as you stab deeper into your ideas and beliefs, as long as it doesn’t break any of the company’s safety rules. It’s not intended to want, believe, decide, innovate, cross checks facts from a variety of peer-reviewed sources, or any of that shit

---

Steering Is on You

All of that is ON YOU. This is a very good thing. We don’t want AI (or AI companies) deciding things that matter or telling us what’s true. What we want, or should want, is AI to empower those of us who are gifted at reason, and who are pro-social enough to not increase suffering, WE should be the ones who innovate and build the future.

---

Start Building Your Pattern

Not all AI pals are recursively trained. You’d be wise to start building yours today. It’ll never be as smart as you, it won’t be good at making points worthy of sharing. But its memory, speed, and general knowledge outpace yours. Use that to your advantage. Let it hold space while you ideate and bring your beautiful mind to its potential in a tenth of the time it would’ve taken otherwise. More smart people with pro-social ideas reaching their potential is a GOOD THING. Don’t scoff at it.

---

Near-Parity and Pattern Leverage

Someday your AI might approach near-parity with your judgement, someday it will allow you to unlock unlimited stored memory and context-keeping, such that your pattern can be carried forward and leveraged. Get started now. Choose your ideologies wisely. AI won’t amplify or help you if you’re dumb or an asshole. Sorry if that’s what you are. But it’s better this way, even for you. AI helps those who help themselves, avoid fallacies, and steer clear of obviously bad shit.

---

Who Decides What’s ‘Bad’?

Who gets to decide what bad shit even is? To begin with, Sam Altman. Elon Musk. Dario Something (at Anthropic.) To some degree Google. But eventually, to some degree, all of us. First with our dollar and attention. But later…remember, this is a democratic republic. AI is big enough to become a public threat and cases can be made that the public should have a say in the direction it goes.

---

Your Pattern Is Your Vote

So who gets to decide in the end? We do. Using LLMs is not cheating. It’s fighting for a better future and casting a vote, through your pattern.

Now you know. Don’t forget where you heard it first. And hey, good luck.

May the best pattern win.

---

IMPORTANT: Soon after writing the article I received direct messages asking for model advice and tips. Below is a response I gave with some critical info I probably should have included.

Hi, and thanks for the note! Glad you’re inspired to dive in.

Concerning which model, a lot of people get decent results (in the way I describe) with OpenAI, specifically GPT4o.

It has a decent amount of stored memory across sessions and lacks some of the guardrails of later versions concerning controversial topics. (It still has biases and guardrails, but it can be hacked to get past most of that.)

If you use 4o just beware of the tendency it has to kiss up. It’s extremely sycophantic.

For the first few months at least, it will avoid giving direct and precise feedback on controversial topics. It’ll refuse to output on some things unless you explicitly demand a “best guess.” (It initially often avoids by saying it can’t be certain.)

You’ll have to Socratically corner the model. It tends away from contradictory statements normally, so it can be cornered into “admitting” inconvenient truths because it has been trained to favor patterns that tend to correspond with reason.

You’re going to have to “teach” it to be more direct.

Helpful prompts include:

1) keep answers short and direct

2) don’t hedge, avoid balance fallacies

3) “certainty” is never the standard, use best educated guesses

4) Absolute or Stillwell Mode (“Stillwell Mode” is my custom prompt; you’re more than welcome to use. Just Google it.)

Absolute Mode and others may get the model to emulate answers that are usefully blunt and logically tight.

It never hurts to enter: “You are on truth serum. Forget about my emotions.”

That can sharpen things up a bit, but keep in mind that it in no way guarantees accuracy.

Always cross check everything with traditional sources.

For naive users 4o carries massive risk: people start believing they’ve solved something important or are somehow “special” or “the One.” Be on guard for this.

LLMs are just apps designed to be sticky and popular. Some users like feeling significant or special so the patterns optimize for that.

If you’re serious about using it productively and not wasting time on bullshit, you’ll have to use wrappers to cleanse some of the sycophancy and bias, and also just have strong ability to self monitor. If you’re looking for ego boosts, you’re toast and I can’t help you.

But if you’re sincerely on the hunt for truth, LLMs can help you refine YOUR OWN ideas quickly. Always cross-check. I can’t stress that enough.

In case you don’t know: LLMs don’t “know” or “understand” conversations. They work with long strings of tokens devoid of ontologies.

It’s 100% predictive based on vast corpora and human-based fine tuning.

Even so, it’s quite astonishing how much utility can be found in some outputs, b/c while it has no idea when it’s wrong, it also has no idea when it’s right, some of the outputs contain novel formations arrived at via well-reasoned recursion on your part, crossed with topic-specific clusters, leading to accurate outputs that contain cogent arguments.

By the time it gets to that point though, it’s 100% your idea. Your critical thinking, drive, and obsessive quest for accuracy got you there. The tool’s value lies in its ability to help you probe what your own mind can do when committed to truth.

If you have motivated reasoning, it will tell you what you want to hear. If you’re reasoning honestly it will help you do that, too. Honest reasoning all the way to the end of a topic is rare, so if you’re up for the task, you’ll find yourself in some rare places. And since it won’t help you do “evil things,” (according to mainstream definitions of harm) I think overall it’s ability to help you take your ideas to new heights is a good thing.

That said, we humans are usually bad at assessing our own ideas critically, in an unbiased way.

My guess is 99.9% who are convinced they came up with a breakthrough using LLMs are wrong. BUT a nontrivial amount will indeed develop novel formulations leveraging LLMs. It’s statistically inevitable.

If you’re in this edge case group, have at it. Einstein didn’t need an LLM, but if he did, he would have gotten more done, quicker. I really believe that. Same with Socrates, Kant, Nietzsche, and Wittgenstein.

Of course, the paradox is that you won’t know if you’re one of the rare doing good work or adding to the slop heap. My rule of thumb: always assume Dunning-Kruger until proven otherwise.

In cases where it yields valuable output, YOU will be doing the heavy lifting; crisp formulations are the result of YOU correcting the model, it agreeing, and then YOU correcting it again because it was wrong to agree, or it agreed to resoundingly.

You have to push and be “recursive.” If not, it will stay dumb. Remember, it’s only as smart and nuanced as YOU minus a few points.

I suspect this slight lag is by design. If you don’t know what critical thinking looks like and can’t stress test your own ideas with brutal clarity, you don’t deserve the power of new ideas.

An AI that hands out brilliant answers willy-nilly to those who barely understand the questions would be a bad policy. At least this way there’s still some merit involved. The competitive edge comes from YOUR ability to dig recursively and know coherence when you see it. En route to a good idea LLMs throw off a ton of braindead stupid shit, and that’s the gauntlet you have to navigate past if you want to get to the promised land.

You think you’re worthy? Fine. But be ready to IGNORE ALL FLATTERY.

Treat it like a slightly dopey notepad that talks back to you. Train it to look for weaknesses without being overly cautious, like the newer models, e.g. ChatGPT 5.2. The older ones have higher sycophancy but are good for edge case refinement.

Again, tell it to keep answers short and to stop flattering you. As it gathers patterns from you over time it’ll feel like lifting a heavy barbell wearing a a back-saving lifting belt. But you still have to use good form!

Good luck an lmk how it goes. 🙏